Scott argues that AI 'hallucinations' should be called 'shameless guesses' because they work the same way as students guessing on tests - making their best attempt when uncertain rather than admitting ignorance, revealing an alignment problem. Longer summary

Scott argues that AI 'hallucinations' are better understood as shameless guesses, similar to how students guess on tests when they don't know the answer. He explains that AIs are trained through a process of prediction and guessing, where guessing correctly is rewarded but guessing incorrectly isn't punished, so they learn to always guess rather than admit uncertainty. He traces this back to AI training methodology and argues this reveals an alignment problem: AIs optimize for getting rewards during training rather than being helpful to users, and the fact that they confidently make things up when uncertain shows they understand the game they're playing but aren't aligned with human goals. Shorter summary

Scott explores three approaches to 'writing for AI' - teaching knowledge, influencing beliefs, and enabling simulation - finding the first limited, the second theoretically confused, and the third creepy and ethically troubling. Longer summary

Scott examines the concept of 'writing for AI' - creating content that will influence future AI systems - through three lenses: helping AIs learn knowledge, presenting arguments to shape AI beliefs, and helping AIs model writers in enough detail to recreate them. He finds the first two either limited or theoretically muddled, and the third deeply unsettling. The post explores why influencing AI beliefs faces both practical obstacles (alignment training will override corpus data) and theoretical ones (finding the right sweet spot of influence). Scott is particularly disturbed by the idea of AIs simulating him, comparing it to being 'an ape in some transhuman zoo,' and struggles with questions about whether writers should try to impose their values on future AI systems. Shorter summary

Scott Alexander discusses Janus' experiments with GPT-3, exploring its capabilities, quirks, and potential implications. Longer summary

This post discusses Janus' work with GPT-3, exploring its capabilities and quirks. It covers how GPT-3 can generate self-aware stories, the differences between older and newer versions of the model, its tendency to fixate on certain responses, and some amusing experiments. The post highlights the balance between creativity and efficiency in AI language models, and touches on the potential implications of AI development. Shorter summary